AI Coding Race: New Models, Tougher Benchmarks, and the Deflating Theorem Economy
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
Kimi K2.7 Code, the coding-focused model from Chinese AI lab Moonshot AI, became generally available inside GitHub Copilot this week, joining a roster that already includes GPT-4o, Claude Sonnet, and Gemini variants. The integration signals that Microsoft, which owns GitHub, is building Copilot as a multi-model interface layer rather than betting exclusively on any single model — a shift in the competitive axis from which company makes the best coding AI to which platform aggregates the best.
ZCode from Z.ai, described as a harness for Zhipu AI's GLM-5.2 model, generated 414 points and 297 comments — among the highest engagement of the day. Community reaction mixed genuine technical curiosity about the harness architecture with ongoing skepticism about benchmark reliability for Chinese frontier models. Several commenters noted that demo quality was high but that independent evaluation lags behind the marketing cycle.
CursorBench 3.1, Cursor's updated evaluation framework, attempts to measure AI coding agents on tasks closer to real engineering work: debugging across file systems, maintaining context in long codebases, handling ambiguous requirements. A related open-source benchmark, Senior SWE-Bench from Snorkel AI, evaluates agents specifically on work expected of a senior engineer — system design tradeoffs, cross-cutting concerns, and underspecified problems. Early scores on Senior SWE-Bench are reportedly significantly lower than on the original SWE-Bench, which the community largely characterized as an honest result.
A philosophical essay titled 'The Fall of the Theorem Economy' by David Bessis provoked a pointed discussion despite a lower score of 74. Its argument: as AI systems demonstrate competence in pattern-matching and empirical validation, the premium on formal theorem-proving as a signal of intelligence is deflating. Mathematicians in the comment thread pushed back sharply, arguing that theorem-proving is not primarily about credentialing — it is the only method available for establishing certain kinds of truth. An AI system correct 99.9% of the time is still wrong about the 0.1%, and for cryptography, formal verification, and safety-critical systems, that gap is not acceptable. The thread's conclusion, broadly, was that formal proof may be declining in prestige for general software development while becoming more essential in safety-critical AI contexts simultaneously.